Impress Your CEO: 5 New AI Strategies for Finance Leaders

The short version

If you've tried ChatGPT, Claude, or Copilot and walked away with a generic answer, you're not alone. The gap between teams running cohort analysis in 2 minutes and teams skipping that analysis entirely keeps growing, and prompting is the bottleneck, not the model.

Nicolas Boucher, founder of the AI Finance Club and a 15-year finance professional turned full-time AI trainer, ran a Ramp webinar on five AI strategies you can use right now. He demonstrated each one live with real datasets and walked through the prompting framework underneath all of them.

Most useful for: FP&A managers, controllers, and CFOs who already use ChatGPT, Claude, or Copilot occasionally and want to go from generic outputs to CFO-ready analysis.

What does AI adoption in finance actually look like right now?

If you've tried AI and walked away frustrated, you're probably treating the model like a search bar. You type a short request, get a generic answer, and conclude the tool is overhyped. As Nicolas explained after thousands of hours training finance teams, the diagnosis is consistent: the tools are capable, but the prompting is wrong.

His fix is to teach repeatable frameworks instead of clever individual prompts. The CSI and FBI framework gives you a checklist for any prompt. Custom GPTs and Claude Projects take it further by encoding that framework once so nobody has to rebuild it.

This matters because your work is repeatable. You write variance commentary every month, build scenario models every quarter, and produce board decks on a calendar. Frameworks compound across those cycles in a way that one-off prompts never will.

5 AI guidelines to follow to get the best output

1. CSI and FBI prompts beat Google-style queries

You probably type AI prompts the way you type Google searches: short, keyword-style, and context-free. Nicolas's CSI and FBI framework fixes that.

CSI stands for Context (who you are), Specific (the situation), and Instruction (what you want). FBI stands for Format (table, bullets, prose), Blueprint (the structure of the output), and Identity (the persona, like a McKinsey consultant).

When you compare a "give me ten ideas to cut cost" prompt against a full CSI and FBI prompt, the difference is obvious. The output goes from generic bullet points to context-aware recommendations you can actually iterate on. Every element you add narrows the model's output toward something useful.

The fastest way to feel the difference is to pick one task you do every week (variance commentary, a budget analysis, a management report). Write a CSI and FBI prompt for it and compare the output against what you'd have produced manually.

"If you use the Google method, like, just saying give me ten ideas to cut cost, you will never get this quality of output. And on top here is only the first draft. After I can iterate, I can choose one of these ten ideas and drill down to get better output."

2. ChatGPT runs price elasticity by handing Excel to Python

The problem you face with analyses like price elasticity isn't always skill, it's setup time. Calculating elasticity in Excel requires at least 3 columns, knowledge of the right formula, and time to get the direction right.

Nicolas demonstrated a different workflow. Upload a simple file with product, volume, and price by month. Then prompt ChatGPT with your role (FP&A manager), the analysis you want (elasticity), and a request to explain the methodology so you can audit it.

ChatGPT doesn't read Excel directly. It hands the file to Python, which reads the first 5 rows, confirms the columns, applies the elasticity formula, and returns the results as a new column.

The result is a complete elasticity analysis across all products and months, faster than building it in Excel, with the formula visible at every step.

The same workflow extends to cohort analysis. Nicolas demonstrated it live from a file with date, customer ID, product, and invoice, and Python produced the visualization in about 20 seconds. His rule: never trust an AI calculation you haven't seen in the console.

"It's also faster than if you want to do it in Excel because in Excel you need at least three columns to do it. You need first the change of price, then the change of volume, and then you need to calculate the relationship between the change of price and the change of volume. Having to do that in Excel and knowing how to do it in the right direction takes longer than using this tool."

3. Claude builds shareable scenario models live

Finance teams often walk into executive meetings with a static three-scenario spreadsheet. When leadership asks "what if we open two more stores instead of one," the team says "let me get back to you," and the conversation ends.

Nicolas built a three-variable model (sales, margin, fixed cost) for a fractional CFO. The client had several stores and wanted to model adding stores, reducing stores, and changes in fixed costs.

The prompt was specific: create a simple dynamic model showing visually how profit changes based on the number of stores, comparing 3 different scenarios, with 3 variables. Claude generated a React-based interface where changing any variable updates the profit projection across all 3 scenarios in real time. The output is a shareable link, not a static spreadsheet.

If you still need the model in Excel, ask Claude for formulas with exact cell references ("use A1, B2, F2") rather than theoretical formulas. Pair that with Excel's Goal Seek, which works backward from a target outcome. In Nicolas's example, Goal Seek determined that hitting a target value of 2,000 required setting the price to 52.

"Now we have this tool where each time you change some parameters, you have the scenario changing in front of your eyes. And you can share this link with everybody. Everybody can use it and run this model by themselves."

4. ChatGPT Canvas turns variance commentary into minutes

ChatGPT Canvas can cut your variance commentary time from hours to minutes. Whether it's budget vs. actuals, MD&A, or board deck narratives, each one requires careful language, specific figures, and a tone that matches your company's voice.

You start with a structured P&L that includes actuals vs. budget, revenue split by country, and cost split by nature. Nicolas prompted ChatGPT using CSI and FBI, specifying his style: professional, bullet-pointed, figures-forward, and actionable. ChatGPT used Python to read the file and then opened a Canvas window, an interactive document editor that rewrites content in real time.

From inside Canvas, you can shorten the commentary, adjust tone, or highlight a specific section and give targeted instructions. Nicolas highlighted the revenue commentary and asked Canvas to use country-level detail to explain where the variance came from. Canvas rewrote just that section, with country-level figures, in seconds.

Nicolas emphasized writing actionable commentary, not descriptive. The goal is a CFO who is in the action, not a CFO who only describes what happened.

"Usually what I do here is professional, bullet points highlighting figures because often it forgets to bring figures if you don't say that. And also actionable because you want to not be a descriptive CFO, but a CFO that is in the action."

5. Custom GPTs encode team style in one shareable link

Custom GPTs solve a problem you've probably already hit: consistency at scale. One person can build the perfect prompt for variance commentary, but when several analysts each tweak it slightly, you get different output styles. Custom GPTs fix that by encoding the prompt, the knowledge, and the style once.

Nicolas opened the backend of a custom GPT he built for financial presentations. It contained his storytelling methodology, slide structure preferences, writing style guidelines, and reference material. When you open that GPT and ask for a revenue forecast slide, the output already reflects every parameter, with no re-prompting needed.

Custom GPTs are the most undervalued feature in ChatGPT today, and almost nobody is using them.

The sharing options are where custom GPTs become useful for your whole team. With ChatGPT Teams or Enterprise, you can share a GPT with everyone in your company, restrict it to internal users, or publish it broadly. Claude Projects offers a parallel capability, and Copilot Studio offers it inside the Microsoft environment. The platform matters less than the act of encoding institutional knowledge once.

"There is something that is quite magical: GPTs. And they are really undervalued. I think almost nobody is using it or using them. And you need to start using GPTs for this."

How do you apply Nicolas's advice this quarter?

Three steps map directly to the strategies above.

1. Rewrite one prompt using CSI and FBI: Pick a task you do weekly (a variance comment, a budget summary, a management report). Write the prompt with all 6 elements, run it, and compare the output against what you'd have produced manually
2. Upload a real dataset and ask for an analysis you've never run before: Try price elasticity or cohort analysis. Pick one your CEO has asked about and you skipped because the Excel setup was too expensive. Hand it to ChatGPT, audit the Python, and bring the result to your next leadership meeting
3. Build one custom GPT for your team and share the link: Encode your reporting style, your preferred commentary structure, and one or two reference documents. Send the link to your direct reports for their next deliverable

None of these takes more than an hour.

Final thoughts

"I really believe in finance. We are one of the well placed to become champions of using AI in the business. And, if people are here, I'm sure they are motivated for that."

Nicolas proved that thesis with real tools, real prompts, and real outputs. The takeaway isn't that AI is the future. It's that the gap between teams who've built repeatable AI workflows and teams who haven't keeps widening.

See how Ramp fits in

These strategies accelerate the analytical and reporting work you do after your data is in place. With Ramp, you get automated expense capture, real-time spend visibility, and AP automation that eliminates the manual work between a transaction and your general ledger. The cleaner the data going into your AI workflows, the better the analysis coming out.

Common questions about AI tooling in finance

Which AI tool should a finance team start with?

The platform matters less than the prompting habit. ChatGPT, Claude, and Copilot all support the CSI and FBI framework. ChatGPT pairs with Python to run calculations and visualizations on uploaded files, and each platform offers a way to encode team style (custom GPTs, Claude Projects, or Copilot Studio). Pick the tool your company already licenses and start there, because switching platforms before you have a repeatable prompt is the wrong order.

How do you keep AI from making errors in financial calculations?

Nicolas's rule is that he won't trust a calculation he hasn't seen the math for. When ChatGPT delegates work to Python, the console shows every step, so read it. For analyses that matter, ask the model to explain its methodology before showing the result, then audit the formula against your own understanding. Treat AI outputs as first drafts that need a controller's eye, not finished deliverables.

Is it safe to upload company financial data to ChatGPT or Claude?

It depends on your license. Corporate and team plans across major providers typically let you opt out of having your data used for training. The underlying security protocols are similar to cloud tools you already use, like Outlook or Google Sheets.

Before uploading a P&L or customer file, confirm your plan's terms and check with your security and IT teams. These strategies assume you're working inside a corporate license.

What is the fastest way to get value out of a custom GPT?

Start with the deliverable your team writes most often. Variance commentary, board narratives, or weekly flash reports are good candidates. Encode the structure, tone, and reference material once, share the link with your direct reports, and ask them to use it for their next round. Iterate on the GPT's instructions based on what they edit out of the output.